InsightFab
Knowledge Base/Common Data Distributions in Manufacturing: Poisson, Binomial, Gamma
Statistical Analysis6 min read

Common Data Distributions in Manufacturing: Poisson, Binomial, Gamma

This article addresses common data reporting challenges in manufacturing, such as unfavorable CPK reports or frequent machine breakdowns, highlighting the critical misconception of assuming a normal distribution for all data. It explains how different data types exhibit distinct distribution patterns, and applying an incorrect analytical method can lead to significant misinterpretations, guiding readers to select appropriate analytical tools based on data characteristics.

That day the CPK report came out, the whole room fell silent for three seconds. Do you know why?

I still remember years ago, when we changed to new materials for our machines, everyone was confident that the yield rate would skyrocket. However, after the first batch of products ran, Quality Control released the CPK report, showing a glaring 1.08. The air in the meeting room instantly froze, and everyone exchanged glances. The boss's face turned green, and he directly asked, "Who can explain to me what kind of distribution this is?" At that time, a rookie engineer foolishly said, "It should be a normal distribution," and was met with a harsh glare from the boss. To be honest, having worked in a semiconductor fab for so long, you always encounter situations like this – yield rate, defects, machine downtime. These data points each have their own unique characteristics.

Where's the Problem?

In essence, what your data "looks like" directly influences how you analyze it and how you identify anomalies. You cannot expect all data to neatly present a beautiful bell curve (which is what we commonly refer to as a normal distribution). Imagine if you were to count the number of machine breakdowns in a day; this is typically an integer with zero decimal places, and most of the time it might be one or two, occasionally five or six. How can you apply a normal distribution to this kind of data?

At this point, we need to understand three extremely common data distributions in manufacturing:

  1. Poisson Distribution: Imagine you are counting "how many times a certain event occurs within a fixed period or space." The most typical examples are "number of machine breakdowns per day," "number of defects on a wafer," or "number of impurity particles found on a production line." The occurrences of these events are independent, the probability of occurrence is very low, but the number of occurrences are countable integers.
  2. Binomial Distribution: This is quite intuitive. It answers the question, "If an event is repeated N times, what is the number of successes (or failures)?" For example, "When producing 1000 chips, what is the number of qualified products?" or "When testing 50 parts, how many are defective?" Each trial has only two outcomes (success or failure), and the probability of success is the same for each trial.
  3. Gamma Distribution: This one is a bit more abstract, but also very useful. It is often used to describe "how long one needs to wait before a certain event occurs." For instance, "the interval time between two machine breakdowns" or "the lifetime of a chip from production to failure." Its values are continuous and usually positive.

So the key is, does your data count "occurrences," "quantities," or "time"? Clarifying this will help you choose the right tool.

How to Do It in Practice?

To determine which distribution to use for your data, there are a few simple criteria:

  1. If you are counting the "number of times an event occurs," and these events have a low probability but many trials, it is usually a Poisson distribution.

* Example: Today you are inspecting a wafer and find 3 defects. Or the DPMO (Defects Per Million Opportunities) is calculated to be 6210. This is typical of Poisson.

* In other words, when you are calculating the number of defects or breakdowns within a certain "unit," and these values are usually small integers, think Poisson.

  1. If you are counting the "number of successes or failures" within a fixed total number of trials, it is usually a Binomial distribution.

* Example: You produced 1000 wafers, of which 980 are good. The "number of good products" here is a result of Binomial.

* In other words, when you have a clear "sample size" N and you count how many individuals "meet a certain condition," then it's Binomial.

  1. If you are measuring the "time interval between two events," or "how long something's lifespan is," it is usually a Gamma distribution.

* Example: Machine A breaks down on average every 8 hours, but it is actually observed to break down sometimes after 5 hours, and sometimes after 12 hours. These breakdown interval times could be a Gamma distribution.

* In other words, when your data is continuous "time" or "value," and this value cannot be negative, and you are interested in situations where an event occurs "after accumulating to a certain extent," then consider Gamma.

Frankly, often you will find data that falls somewhat in between. In this case, you can try running a Goodness-of-Fit Test with statistical software to see which distribution best represents your data.

The Most Common Pitfall

The biggest pitfall I've encountered is that many people habitually analyze all data as if it were a normal distribution. The result is that the CPK looks great, but the yield rate just won't improve.

I remember once, when we were analyzing the particle count for a certain process, an engineer directly calculated the mean and standard deviation, and then used a normal distribution to draw a control chart. As a result, several points on the chart consistently exceeded the limits, but no one could find the reason. Later, when I looked at the data, I found that most of the time the particle count was 0 or 1, occasionally jumping to 5 or 6. This was not a normal distribution at all! It was a typical Poisson distribution. When we switched to a control chart suitable for Poisson distribution to analyze it, those "abnormal points" truly revealed their underlying meaning, and we finally identified that the issue was due to an overly long consumable replacement cycle. To be honest, at that time, I truly felt as if I had been "deceived" by the data.

One Thing You Can Do Today

Open your latest yield report, look at the numbers, and try to determine if they are typical of Poisson, Binomial, or Gamma distributions.

Want to try it yourself?

Every tool mentioned in this article is available on InsightFab — just upload a CSV to analyze.

Go to Tools

More in this category

Correlation Coefficient: Pearson vs Spearman – Linear vs Monotonic Relationships

This article sheds light on a common pitfall in engineering data analysis: the misapplication of correlation coefficients. It explains why relying solely on Pearson correlation, which primarily detects linear relationships, can lead to incorrect conclusions when data exhibits non-linear, monotonic trends, especially when addressing issues like unexpected product yield drops.

Post-hoc Test: Choosing Between Tukey vs Bonferroni

This article addresses a common production line challenge: identifying the specific source of a problem when overall metrics like CPK fall short. It introduces Post-hoc Tests as a crucial tool to pinpoint significant differences between individual machines or groups, enabling precise problem-solving and immediate action.

Statistical Power: How to Ensure Your Conclusions Are Meaningful

This article addresses a common challenge for engineers: drawing conclusions solely based on observed data changes without accounting for statistical significance. It emphasizes the importance of using 'Statistical Power' to validate findings, ensuring that conclusions are robust and well-supported, particularly when presenting to management.